Large-Scale Fluorescence Calcium-Imaging Methods for Studies of Long-Term Memory in Behaving Mammals

Abstract

During long-term memory formation, cellular and molecular processes reshape how individual neurons respond to specific patterns of synaptic input. It remains poorly understood how such changes impact information processing across networks of mammalian neurons. To observe how networks encode, store, and retrieve information, neuroscientists must track the dynamics of large ensembles of individual cells in behaving animals, over timescales commensurate with long-term memory. Fluorescence Ca²⁺-imaging techniques can monitor hundreds of neurons in behaving mice, opening exciting avenues for studies of learning and memory at the network level. Genetically encoded Ca²⁺ indicators allow neurons to be targeted by genetic type or connectivity. Chronic animal preparations permit repeated imaging of neural Ca²⁺ dynamics over multiple weeks. Together, these capabilities should enable unprecedented analyses of how ensemble neural codes evolve throughout memory processing and provide new insights into how memories are organized in the brain.

In vivo, calcium-imaging methods offer insight into how neural networks encode, store, and retrieve information. These studies depend on the use of appropriate fluorescent indicators, cell-targeting methods, and imaging systems.

Recent years have brought major new capabilities for manipulating mammalian neural circuits involved in memory processing. Advances in genetics, viral delivery methods, and optogenetic and pharmacogenetic techniques have allowed researchers to explore how specific cell types and neural projection pathways contribute to memory processing in behaving rodents (Ciocchi et al. 2010; Goshen et al. 2011; Letzkus et al. 2011; Garner et al. 2012; Liu et al. 2012; Nguyen-Vu et al. 2013; Cowansage et al. 2014; Redondo et al. 2014; Senn et al. 2014; Wolff et al. 2014). By using optogenetic or pharmacogenetic modes of neural excitation in associative learning paradigms, one can substitute for conditioned (Choi et al. 2011; Kwon et al. 2014) or unconditioned (Johansen et al. 2010; Kimpo et al. 2014) stimuli, create synthetic memory traces (Garner et al. 2012), artificially evoke previously learned behaviors (Liu et al. 2012; Cowansage et al. 2014; Kim et al. 2014), and even condition mice to associate pairs of stimuli that were never actually presented together (Ramirez et al. 2013; Redondo et al. 2014). One optogenetic study comparing the behavioral effects of fast neural inhibition against those of inactivation over longer time-scales has also demonstrated that the set of brain circuits required for memory function depends strongly on the timescale at which one probes this necessity (Goshen et al. 2011), likely because slower methods of inactivation allow the brain to attain partial or total recoveries of function. Overall, recent advances in the causal manipulation of mammalian neural circuits have yielded important gains in the delineation of necessary and sufficient neural events for memory performance.

To complement the new approaches for perturbing circuits, neuroscientists also need improved observational methods for probing the neural representations the mammalian brain uses normally for memory processing and storage. In many cases, these representations appear to be distributed over large networks of cells and multiple brain areas (Tse et al. 2007; Goshen et al. 2011; Cowansage et al. 2014; Redondo et al. 2014). As computer scientists well appreciate, the manner in which information is represented in storage can have a major impact on the ease of retrieving information, encoding objects’ attributes along with their identities, classifying and grouping items according to their characteristics, and associating items that have defined relationships such as in space or time (Date 2000). To understand how the brain accomplishes these feats, we need in vivo recording methods capable of revealing large-scale neural representations and tracking their evolution over timescales pertinent to long-term memory. Such methods would open the door to sophisticated studies of how large-scale neural dynamics and codes may facilitate or hinder different forms of information management in the mammalian brain (Sadtler et al. 2014).

Large-scale Ca²⁺-imaging techniques offer considerable promise toward achieving studies of this kind (Dombeck et al. 2010; Komiyama et al. 2010; Harvey et al. 2012; Huber et al. 2012; Ziv et al. 2013; Heys et al. 2014; Low et al. 2014; Modi et al. 2014; Peters et al. 2014; Rickgauer et al. 2014; Hamel et al. 2015). Key technical advantages include the capacity to monitor the dynamics of hundreds of cells concurrently, target specific neuron types for study based on their genetic identities or connectivity patterns, reliably track individual cells for many weeks in behaving animals, extract the signals of individual neurons nearly regardless of their activity rates, and visualize the anatomical organization of memory storage at the cellular scale. Optical methods for Ca²⁺ imaging also lend themselves naturally to combined usages with other optical methods (Prakash et al. 2012; Deisseroth and Schnitzer 2013; Jorgenson et al. 2015), such as optogenetics (Rickgauer et al. 2014; Szabo et al. 2014; Grosenick et al. 2015; Packer et al. 2015), fluorescent tagging of activated neurons (Reijmers et al. 2007; Liu et al. 2012; Guenthner et al. 2013; Kawashima et al. 2013; Ramirez et al. 2013; Redondo et al. 2014), and postmortem imaging of optically cleared tissues (Hama et al. 2011; Chung et al. 2013; Ke et al. 2013; Susaki et al. 2014; Yang et al. 2014a). In the future, integrated optical studies will likely combine observations of neurons’ normal activity patterns, precise optogenetic manipulations of these dynamics, and detailed postmortem examinations of the dendritic morphologies, axonal projections, and macromolecular architectures of the very same cells (Deisseroth and Schnitzer 2013; Jorgenson et al. 2015).

This review aims to introduce in vivo Ca²⁺-imaging methods to researchers studying mammalian learning and memory. The next sections provide technical information for practitioners of in vivo Ca²⁺ imaging, emphasizing methodological strengths and limitations. Later sections discuss the types of experiments on learning and memory that Ca²⁺-imaging techniques are likely to enable in the near future.

Ca²⁺ IMAGING AS A MEANS OF INFERRING NEURONAL SPIKING DYNAMICS

Nearly all neuron types express voltage-sensitive Ca²⁺ channels (Trimmer and Rhodes 2004). Fluorescence Ca²⁺ imaging as a means of detecting somatic spiking dynamics relies on increases in intracellular Ca²⁺ ion concentration, [Ca²⁺], which occur in response to neural membrane depolarization (Helmchen et al. 1996; Grienberger and Konnerth 2012; Hamel et al. 2015). An essential part of all Ca²⁺-imaging methodologies is the Ca²⁺ indicator, a fluorescent reporter molecule that changes its photophysical properties in response to variations in intracellular [Ca²⁺]. The detailed biophysical processes underlying Ca²⁺ imaging have been reviewed at length elsewhere (Grienberger and Konnerth 2012). Thus, this section focuses more on the enabling features and pitfalls that memory researchers should consider when contemplating the use of a Ca²⁺ indicator as a means of visualizing spiking activity.

A fluorescent Ca²⁺ indicator molecule can bind one or more Ca²⁺ ions, which thereby alters the indicator's optical properties (Tian et al. 2009; Grienberger and Konnerth 2012). In general, there are distinct classes of indicators that undergo different types of photophysical changes, such as in fluorescence intensity, absorption or emission spectra, or emission lifetime (Grienberger and Konnerth 2012). For studies of long-term memory, researchers will generally want to choose a Ca²⁺ indicator that is genetically encoded (Mank et al. 2008; Grienberger and Konnerth 2012; Looger and Griesbeck 2012; Tian et al. 2012) owing to the capacity for stably expressing these protein sensors over time courses commensurate with those of long-term memory storage, such as for the types of experiments considered here in later sections.

For Ca²⁺-imaging studies in behaving mammals, the genetically encoded indicator that is most widely used at present (e.g., based on gene requests to Addgene) is GCaMP6, which can bind up to four Ca²⁺ ions and reports [Ca²⁺] rises via increases in the intensity of its green fluorescence emissions (Chen et al. 2013). GCaMP6 has a greater dynamic range of fluorescence signaling than prior genetically encoded Ca²⁺ indicators. It also has Ca²⁺ binding kinetics rivaling those of the best synthetic Ca²⁺ indicators used for acute in vivo imaging studies (Garaschuk et al. 2006, Komiyama et al. 2010, Modi et al. 2014). Emerging types of genetically encoded red fluorescent Ca²⁺ indicators with spectrally distinct emissions from GCaMP6 open up interesting possibilities for monitoring the dynamics of two distinct cell types simultaneously (Inoue et al. 2015).

Despite the impressive capabilities of GCaMP6, no Ca²⁺ indicator provides an exact readout of membrane voltage dynamics (Fig. 1A–H). The biophysical processes that govern the [Ca²⁺] rise in response to an action potential include Ca²⁺ channel activation and intracellular Ca²⁺ buffering and are distinct from those setting the action potential's electrical waveform (Helmchen et al. 1996). The time-dependence of the signals from a Ca²⁺ indicator also depends strongly on its Ca²⁺ handling properties, particularly the kinetic rates, equilibrium constant and Hill coefficient for Ca²⁺ binding and unbinding (Fig. 1A–D,G,H) (Sun et al. 2013). Hence, beyond the intrinsic differences between the dynamics of transmembrane voltage and intracellular Ca²⁺ excitation, the signals from a fluorescent Ca²⁺ indicator represent a temporally filtered version of the intracellular [Ca²⁺] time course. Because of the experimental difficulty of directly relating a cell's time-varying fluorescence emissions to its intracellular [Ca²⁺] dynamics in a live brain, virtually all researchers express fluorescence time traces from in vivo Ca²⁺-imaging studies in units of a percentage change from each cell's baseline fluorescence level, typically denoted ΔF(t)/F₀.

Figure 1.

The GCaMP6 family of genetically encoded Ca²⁺ indicators allows high-fidelity detection of action potentials from neocortical pyramidal cells in live mice and ∼20–200 msec spike timing estimation accuracy. (A) In pyramidal neurons expressing GCaMP6s (black traces) or GCaMP6f (cyan traces) in neocortical visual area V1 of live mice, simultaneous loose-seal cell-attached electrical and two-photon fluorescence Ca²⁺-imaging recordings show the reliability of the optical response to each action potential. Asterisks below the electrical traces (lower traces in each pair) mark occurrences of individual action potentials; numerals mark action potential bursts and report the number of spikes in each burst. (Inset) Two-photon image of a neuron expressing GCaMP6s; the recording pipette is indicated by the red lines. Scale bar, 10 µm. (B) Expanded view of the fluorescence traces in A over the time periods delineated by the dashed boxes. Dashed vertical lines mark action potential occurrences. (C) Fluorescence changes in response to one action potential for GCaMP6s (upper) and GCaMP6f (lower). Gray traces show the optical responses to individual spikes. Black and cyan traces show the average response. (D) Traces of the median fluorescence change in response to one action potential for the genetically encoded Ca²⁺ indicators GCaMP5K (green trace), GCaMP6f (cyan), GCaMP6m (magenta), and GCaMP6s (black). Shading indicates s.e.m. (E) Peak fluorescence changes (mean ± s.e.m.) observed by Ca²⁺ imaging as a function of the number of spikes detected electrically in a 250 msec time bin during simultaneously acquired optical and electrical recordings in V1 neurons in live mice. Color key is the same as in D. (F–H) Comparative performance metrics of the different indicators. (F) Percentage of action potentials correctly detected when the detection threshold is set to yield only 1% false positives. (G) Time constant for the decay of fluorescence following an action potential occurrence. (H) Time constant for the fluorescence rise in response to an action potential. All error bars are s.e.m. All panels derived and modified from Chen et al. (2013).

There are three main variants of GCaMP6 (slow, medium, and fast) that have different Ca²⁺-binding kinetics and, hence, temporally filter intracellular [Ca²⁺] dynamics to different degrees (Fig. 1) (Chen et al. 2013). The slow variant, GCaMP6s, binds Ca²⁺ more tightly than the fast variant, GCaMP6f. This difference confers greater Ca²⁺ sensitivity to GCaMP6s and leads to fluorescence transients of longer duration (Fig. 1C), but sacrifices signaling speed and the ability to distinguish action potentials occurring in quick succession (Fig. 1A,G,H). For GCaMP6s, the fluorescence signals in response to an action potential have a rise time-constant of ∼180 msec and a decay time constant of ∼580 msec, whereas, for GCaMP6f, these values are ∼30 msec and ∼150 msec, respectively (Fig. 1G,H) (Chen et al. 2013). Cooperation and saturation effects in the binding of multiple Ca²⁺ ions to a GCaMP6 molecule also cause nonlinearities in the temporal filtering.

In addition to these aspects of indicator dynamics, stochastic fluctuations in the emission and detection of fluorescence photons constrain the fidelity of action potential detection and the accuracy of spike-timing estimation (Wilt et al. 2013; Hamel et al. 2015). Photon “shot noise” has its origins in quantum mechanics and sets physical limits on spike detection fidelity and timing accuracy—even when additional noise sources, such as instrumentation noise, are minuscule. Signal detection and estimation theories provide theoretical lower bounds on the spike detection and timing errors incurred because of fluctuations in photon counts (Wilt et al. 2013; Hamel et al. 2015).

Because of these biophysical and optical facets of Ca²⁺ imaging, there is substantial variability in the relationship between the amplitude and waveform of a somatic Ca²⁺ transient, as seen by fluorescence imaging, and the number of action potentials underlying the transient (Fig. 1C). Moreover, the relationship between membrane voltage and somatic [Ca²⁺] dynamics differs between different types of neurons. Cell types vary regarding the expression and properties of their Ca²⁺ channels, levels of intracellular [Ca²⁺] buffering, and other aspects of Ca²⁺ signaling such as Ca²⁺-induced Ca²⁺ release from intracellular stores. Whenever feasible, it is best to empirically determine the relationship between Ca²⁺-related fluorescence signals and the underlying spike trains in the specific cell type of interest. A common approach is to perform simultaneous electrical recordings and fluorescence measurements in the exact same cells, such as in a brain slice (Tian et al. 2009), or by using in vivo Ca²⁺ imaging to guide the targeting of an electrode to a selected cell in a live animal (Fig. 1) (Chen et al. 2013).

Simultaneous optical and electrical recordings from visual cortical pyramidal neurons in live mice have revealed approximately linear relationships between the peak amplitude of a somatic Ca²⁺ transient and the underlying number of action potentials (Fig. 1E) (Chen et al. 2013). Other types of pyramidal cells likely also exhibit similar linear relationships between these parameters, but not necessarily with the same linear function. More generally, in neurons with temporally sparse spiking patterns, in vivo Ca²⁺ imaging usually reveals discrete Ca²⁺ transients representing one or more spikes.

Inhibitory interneurons have distinct Ca²⁺ signaling and buffering attributes from those of excitatory neurons, and they often fire spikes at fast rates or in bursts. In such fast-spiking neurons, it is typically infeasible to assign precise spike numbers to individual Ca²⁺ transients. Still, Ca²⁺ imaging generally reveals the overall patterns of how interneurons modulate their time-dependent activity rates; for instance, recent Ca²⁺-imaging studies of visual cortical interneurons have monitored visually evoked increases in somatic fluorescence intensity, rather than incidences of individual Ca²⁺ transients (Kerlin et al. 2010; Runyan et al. 2010; El-Boustani and Sur 2014).

Regardless of cell type, in live or behaving animals, there are yet additional noise sources that can make it challenging to infer the exact number of spikes underlying any Ca²⁺ transient. These noise sources are generally nonstationary over time and can originate from hemodynamic events, brain motion artifacts induced by physiological rhythms or voluntary movements, or fluorescence Ca²⁺ signals from neuropil activation. Notwithstanding, there is almost always a useful, monotonic relationship between the mean intensity of somatic fluorescence emissions from a Ca²⁺ indicator and the rate of somatic spiking.

In vivo Ca²⁺ imaging generally does not reliably reveal membrane hyperpolarization or other subthreshold aspects of somatic voltage dynamics. However, outside the soma, by sparsely expressing a Ca²⁺ indicator in isolated cells, one can detect in vivo the subcellular Ca²⁺ activation patterns of individual dendrites, dendritic spines, and even axons (Bock et al. 2011; Petreanu et al. 2012; Xu et al. 2012; Chen et al. 2013; Glickfeld et al. 2013; Kaifosh et al. 2013; Lovett-Barron et al. 2014; Sheffield and Dombeck 2015).

Overall, although in vivo Ca²⁺ imaging may not be able to adjudicate issues that hinge on exact spike counts and millisecond-scale timing, there are many questions about neural coding and memory processing that Ca²⁺ imaging can answer. The latter questions might focus on the identities of cells that encode specific types of information, the acquisition and long-term stability of coding properties under different behavioral conditions, the relationships between coding properties and other cellular attributes, the extent to which ensembles of neurons encode information in a cooperative manner, how neural codes change across different brain states, or the anatomical relationships between cells involved in storing specific memories.

TARGETING SPECIFIC SUBSETS OF CELLS FOR EXPRESSION OF A GENETICALLY ENCODED Ca²⁺ INDICATOR

To study large-scale neural coding by Ca²⁺ imaging, it is important to consider the various ways in which one can express a genetically encoded indicator in the cells of interest. To date, most Ca²⁺-imaging studies in behaving mammals based on a genetically encoded Ca²⁺ indicator have used an adeno-associated viral (AAV) vector to target indicator expression to specific neuron types at the virus injection site (Dombeck et al. 2010; Komiyama et al. 2010; Harvey et al. 2012; Huber et al. 2012; Ziv et al. 2013). A key consideration with the use of an AAV concerns the viral serotype, because distinct serotypes of AAV enter different neuron types to varying degrees (Aschauer et al. 2013). Another key factor is the genetic promoter used to express the Ca²⁺ indicator. For instance, an AAV whose genome incorporates a CaMK2a promoter can target indicator expression to the specific subset of virally infected neurons in which this promoter is active, such as pyramidal neurons (Ziv et al. 2013). A pan-neuronal promoter, such as the Synapsin promoter, drives expression in nearly all neuron types (Dombeck et al. 2010; Huber et al. 2012). However, many promoters of interest are too large to be packaged within an AAV. There are other viruses capable of delivering larger genetic payloads, but they tend to be either toxic to neurons or less effective at expressing a fluorescent indicator at the levels needed for adequate brightness.

Thus, a widely used alternate means of targeting specific cell types is to create an AAV that selectively drives indicator expression over the long-term in cells expressing the enzyme Cre-recombinase (Sauer 1998). Many different Cre-driver mouse lines, transgenic animals that express this enzyme in particular classes of cells, are widely available (Madisen et al. 2010, 2015). Hence, a single viral construct that expresses the Ca²⁺ indicator in a Cre-dependent manner can be fruitfully combined with a wide variety of Cre-driver lines. This modular approach is generally more efficient and feasible than creating a separate virus for each cell type of interest.

In addition to facilitating studies of somatic Ca²⁺ activity, AAV-based strategies for GCaMP expression have enabled observations of Ca²⁺ activity in specific neural pathways, by imaging the Ca²⁺ dynamics of axons of neurons whose cell bodies lie elsewhere (e.g., in a brain area distal to the imaging field of view). This approach selectively confines the Ca²⁺ signals to neurons that have their cell bodies at the viral injection site and project axons to the imaging site (Petreanu et al. 2012; Xu et al. 2012; Glickfeld et al. 2013; Kaifosh et al. 2013; Lovett-Barron et al. 2014). For example, in higher visual areas of the mouse neocortex, this approach allowed selective imaging of axons from neurons with cell bodies in visual area V1 (Glickfeld et al. 2013). A promising alternative to this approach for imaging pathway specific Ca²⁺ activity is to use canine adenovirus-2 (CAV-2), which efficiently infects neuronal axons (Hnasko et al. 2006; Bru et al. 2010; Boender et al. 2014).

Other emerging strategies for expressing a genetically encoded Ca²⁺ indicator include the use of transgenic mice (Chen et al. 2012b; Zariwala et al. 2012; Dana et al. 2014; Madisen et al. 2015) or transsynaptic viruses (Osakada et al. 2011). Until recently, transgenic mice created to express genetically encoded Ca²⁺ indicators usually suffered from dimmer fluorescence levels than those attainable by viral delivery methods. This is because transgenic mice typically had one, or only a few, copies of the Ca²⁺ indicator gene. By comparison, many AAV particles can enter an individual infected cell, which boosts each cell's expression of the protein fluorophore. This issue is now widely appreciated, and a promising transgenic mouse strategy that uses a triple transgene approach to increase Ca²⁺ indicator expression levels has recently emerged (Madisen et al. 2015).

Transsynaptic neurotropic viruses that traverse neuronal synapses also offer interesting prospects for targeting Ca²⁺-imaging studies to cells receiving or providing specific synaptic inputs. Retrograde transsynaptic viruses include rabies (Kelly and Strick 2000; Taber et al. 2005; Wickersham et al. 2007a,b; Miyamichi et al. 2011; Osakada et al. 2011; Osakada and Callaway 2013) and pseudorabies (Enquist et al. 2002; Card et al. 2011; Card and Enquist 2014; Oyibo et al. 2014). The use of these viruses for anatomical mapping studies is established but applications to Ca²⁺ imaging remain somewhat experimental (Osakada et al. 2011). A key advantage of rabies has been that genetically modified versions of rabies exist that can pass retrograde across a single synapse, but not across disynaptic or polysynaptic projections (Etessami et al. 2000; Wickersham et al. 2007a,b). This restriction confines the set of cells under study to those with well-defined monosynaptic connections. Analogous versions of pseudorabies are emerging (Oyibo et al. 2014). Transsynaptic labeling approaches based on anterograde viruses remain less well developed. Anterograde labeling methods exist based on the herpes simplex virus but are highly toxic for the infected neurons (Hoover and Strick 1999; Lo and Anderson 2011).

Another strategy for studying specific cells involved in memory processing is to target cells that undergo immediate early gene (IEG) activation at specific phases of a behavioral protocol. Memory researchers often use IEG expression as a neuronal marker of engagement in memory formation, such as via spiking or synaptic plasticity, depending on the particular IEG (Guzowski and Worley 2001; Czajkowski et al. 2014). There are several different types of transgenic mice that permit fluorescence tagging of specific cells according to their patterns of Fos or Arc gene activation (Reijmers et al. 2007; Garner et al. 2012; Liu et al. 2012; Guenthner et al. 2013; Ramirez et al. 2013; Czajkowski et al. 2014; Redondo et al. 2014). Beyond readout of cell activation, IEG-based tagging has chiefly been used in combination with pharmacogenetic or optogenetic means of selectively reactivating or silencing tagged neurons to probe their causal roles in memory processing (Garner et al. 2012; Liu et al. 2012; Ramirez et al. 2013; Cowansage et al. 2014; Redondo et al. 2014; Yiu et al. 2014).

Similar strategies involving IEG-based tagging appear promising for Ca²⁺-imaging studies. For instance, one might selectively image Ca²⁺ activity in neurons that underwent IEG activation at an earlier timepoint in a behavioral protocol, to probe how the activity of neurons involved in memory encoding (based on IEG expression) relates to their spiking dynamics at memory recall. More broadly, joint monitoring of IEG expression and Ca²⁺ imaging might allow comparisons of how information is represented across different sets of neurons, such as those in which Fos, Arc, or other IEGs are activated on different days in a long-term experiment. This type of study might help adjudicate recent hypotheses proposing that the activation of certain IEGs may prime neurons for upcoming bouts of memory storage—before the remembered events have even occurred (Han et al. 2007; Rogerson et al. 2014). Such issues might be fruitfully studied through the use of genetic constructs that express a pair of differently colored fluorescence indicators, one reporting the level of IEG activation and the other signaling Ca²⁺ activity (Kawashima et al. 2013).

OPTICAL INSTRUMENTATION FOR IN VIVO Ca²⁺ IMAGING IN AWAKE BEHAVING ANIMALS

Ca²⁺-imaging studies in behaving rodents generally use one of two main types of optical instrumentation. In some studies, the animal is head-fixed and behaves in place under the objective lens of a conventional upright two-photon fluorescence microscope (Dombeck et al. 2007; Nimmerjahn et al. 2009). In other studies, the animal carries a miniature fluorescence microscope on its head, thereby allowing Ca²⁺-imaging studies of neuronal activity during unconstrained animal behavior (Helmchen et al. 2001; Flusberg et al. 2008). Both approaches allow long-term imaging over weeks and have enabled Ca²⁺-imaging studies of the neurobiology of memory (Komiyama et al. 2010; Harvey et al. 2012; Huber et al. 2012; Ziv et al. 2013; Peters et al. 2014). For a comparison of the optical issues arising in the two imaging formats and discussion of the image analysis algorithms used to extract individual cells’ time-varying Ca²⁺ signals from the fluorescence videos, we refer readers to a recent review of these engineering matters (Hamel et al. 2015). Here we focus on considerations important to the design of scientific studies.

A key consideration is whether limiting an animal's range of behavior via head fixation will be a benefit or a drawback. To address many scientific questions, it is crucial to gather data from a large set of stereotyped trials, and the behavioral constraints imposed by head fixation can facilitate both the controlled delivery of sensory stimuli (Verhagen et al. 2007; Carey et al. 2009; Andermann et al. 2011; Blauvelt et al. 2013; Patterson et al. 2013; Miller et al. 2014) and behavioral stereotypy in an animal's responses (Huber et al. 2012; Masamizu et al. 2014; Peters et al. 2014). Together, these capabilities have allowed Ca²⁺-imaging studies in head-fixed animals trained to perform perceptual discrimination (Andermann et al. 2010; Komiyama et al. 2010; O’Connor et al. 2010) and motor execution tasks (Huber et al. 2012; Masamizu et al. 2014; Peters et al. 2014). A recent Ca²⁺-imaging study even examined a version of associative fear conditioning adapted for head-restrained animals (Lovett-Barron et al. 2014).

In addition to its compatibility with conventional two-photon microscopy, the head-restrained imaging format is also well suited for use with custom-designed fluorescence microscopy setups that provide novel imaging capabilities (Horton et al. 2013; Heys et al. 2014; Lecoq et al. 2014; Low et al. 2014; Quirin et al. 2014; Stirman et al. 2014; Bouchard et al. 2015) or combine two-photon Ca²⁺ imaging and two-photon optogenetic capabilities in behaving mammals (Packer et al. 2012; Prakash et al. 2012; Rickgauer et al. 2014; Packer et al. 2015). As new optical brain-imaging modalities emerge, the steadily increasing number of behavioral assays for head-fixed rodents will provide valuable test beds for validating novel optical approaches and then capitalizing on the resulting opportunities for biological experimentation.

Head-fixation can also be combined with approaches in which an animal navigates a virtual reality while walking or running in place (e.g., on a spherical treadmill) (Fig. 2) (Holscher et al. 2005; Harvey et al. 2009; Dombeck et al. 2010; Keller et al. 2012; Ravassard et al. 2013; Sofroniew et al. 2014; Aghajan et al. 2015). Virtual reality methods permit sensory manipulations that would be difficult or impossible to achieve in a freely behaving animal (Keller et al. 2012). To date, most virtual reality systems for rodents have involved projection of a visual scene around the head-fixed animal, such as by using a toroidal screen (Fig. 2A,B) (Holscher et al. 2005; Harvey et al. 2009; Dombeck et al. 2010), or arrays of two or more flat video monitors (Fig. 2C) (Keller et al. 2012). However, virtual reality methods that provide tactile feedback are now also emerging (Sofroniew et al. 2014).

Figure 2.

Apparatus for projecting a virtual reality environment. (A) Schematic of a virtual reality apparatus. An image of a virtual space is projected by a digital light-processing (DLP) projector, deflects off a reflecting mirror (RM), and is magnified by a curved, angular amplification mirror (AAM) onto a toroidal screen. A head-fixed rodent is at liberty to walk or run in place on an air-suspended spherical treadmill. An optical computer mouse monitors the treadmill's rotations in both angular dimensions (modified from Harvey et al. 2009; see also Figures 1 and 2 in Holscher et al. 2005). (B,C) Examples of commercially available apparatus, using a toroidal screen, B, and a set of six computer screens, C, that create virtual environments in which head-fixed rodents can navigate. (Images obtained from and used with permission of PhenoSys GmbH.)

A noteworthy caveat is that in an animal exploring a virtual environment, neural dynamics may deviate from their normal forms during unrestrained animal behavior (Ravassard et al. 2013), perhaps because of altered vestibular, self-motion, or visual or tactile sensory inputs. The extent of these effects remains unclear, especially for long-term experiments in which neural plastic effects might accrue over time, and the degree to which different forms of virtual reality impact neural coding remains under active investigation (Ravassard et al. 2013; Aronov and Tank 2014; Aghajan et al. 2015). As computer graphics and reality simulation methods advance, neuroscientists will gain increasingly sophisticated capabilities for performing sensory manipulations during large-scale Ca²⁺ imaging. Notably, virtual reality methods have been used to study spatial memory in humans in subjects undergoing functional magnetic resonance brain imaging (Pine et al. 2002) or intracranial electrical recordings (Suthana et al. 2012). This commonality might enable parallel studies in which animal and human subjects perform similar tasks, but during different forms of brain imaging suitable for each species.

Complementary to Ca²⁺-imaging techniques that require head fixation of an alert animal, miniature head-mounted fluorescence microscopes that have flexible cables (optical fibers or floppy electrical lines) allow Ca²⁺-imaging studies in freely behaving mammals (Helmchen et al. 2001; Flusberg et al. 2008; Sawinski et al. 2009; Ghosh et al. 2011; Ziv et al. 2013; Berdyyeva et al. 2014; Betley et al. 2015; Jennings et al. 2015). The ability to study unrestrained forms of animal behavior is an important capacity, because many animal behaviors are incompatible with or poorly suited to study under conditions of head restraint. Examples include the social behaviors such as fighting, mating, care giving, and other forms of interaction; behaviors probing fear, stress, or anxiety in which head restraint may cause confounding behavioral or physiological effects; and many motor and vestibular-dependent behaviors that involve self-motion cues. All of these behaviors can have learning and memory components. Miniaturized microscopes are also compatible with most of the behavioral assays that are already widely deployed and validated across neuroscience research institutions and in the pharmaceutical industry.

At present, the most widely used miniature fluorescence microscope in the neuroscience research field is a two-gram integrated microscope for Ca²⁺ imaging in freely behaving mice (Fig. 3A–C) (Ghosh et al. 2011; Ziv et al. 2013). This device is optically integrated in the sense that each two-gram unit contains all required optical components, including a light-emitting diode (LED) that provides illumination, miniature lenses, a tiny fluorescence filter cube, and a cell phone camera chip that captures the fluorescence images (Fig. 3A,B). Fine electrical wires carry power and control signals to the microscope and convey digital images from the camera chip to an external data-acquisition box. The integrated microscope is commercially available and generally compatible with the spatial mazes, operant chambers, and fear-conditioning boxes commonly used for behavioral studies in mice. With multiple integrated microscopes, one can perform fluorescence Ca²⁺ imaging in several freely behaving mice in parallel.

Figure 3.

The miniature integrated fluorescence microscope for Ca²⁺ imaging in freely behaving mice. (A) Computer-assisted design of an integrated fluorescence microscope, showing the optical and mechanical components in cross section. (B) Photograph of an assembled microscope. Insets show the printed circuit boards holding the cell phone camera chip (upper right), the blue light-emitting diode (lower right), and the fluorescence filter cube (lower left). (C) Integrated microscope mounted on a freely behaving mouse. (Inset) An integrated microscope on the tip of a finger. Scale bars, 5 mm (A and B) and apply also to the insets. Panels A and B are from Ghosh et al. (2011). (Images in panel C are courtesy of Inscopix.)

When used in conjunction with microendoscope probes (350–1000 µm diameters) that can be permanently implanted deep in brain tissue (Jung and Schnitzer 2003; Jung et al. 2004; Levene et al. 2004; Barretto et al. 2011), the integrated microscope allows brain imaging in a wide variety of different brain areas. These include neocortex, hippocampus, striatum, amygdala, hypothalamus, nucleus accumbens, substantia nigra, and cerebellar cortex (Ghosh et al. 2011; Ziv et al. 2013; Grewe et al. 2014; Lefort et al. 2014; Parker et al. 2014; Betley et al. 2015; Hamel et al. 2015; Jennings et al. 2015). To monitor the Ca²⁺ dynamics of large sets of individual cells in these brain areas across multiple weeks, researchers can repeatedly image the same field of view simply by reattaching the integrated microscope to a permanently affixed cranial base plate at the start of each imaging session (Ziv et al. 2013). Electroencephalography or electromyography recordings can also be performed simultaneously with Ca²⁺ imaging, such as for monitoring brain or locomotor states (Berdyyeva et al. 2014). A single field of view (∼0.5–0.6 mm²) from the integrated microscope can routinely reveal many hundreds of neurons over the course of an extended Ca²⁺-imaging study (Ziv et al. 2013), and sometimes >1000 cells (Fig. 4) (Alivisatos et al. 2013).

Figure 4.

A Ca²⁺-imaging dataset from a single freely behaving mouse can contain ∼1000 neurons. (A) 1202 GCaMP-expressing CA1 hippocampal pyramidal neurons (red somata), identified in Ca²⁺-imaging data taken in a freely moving mouse using the integrated microscope, shown atop a mean fluorescence image (green) of CA1. Vessels appear as dark shadows (from Alivisatos et al. 2013). (B) 35 example traces of Ca²⁺ activity from CA1 hippocampal pyramidal neurons expressing GCaMP3 under the control of the Camk2a promoter (modified from supplementary Fig. 1b in Ziv et al. 2013).

We expect that the optical methods for Ca²⁺ imaging in head-restrained and freely behaving animals will both continue to advance in their capabilities. Likely improvements include optical access to greater numbers of cells per animal, fast volumetric imaging, multicolored Ca²⁺ imaging for monitoring two targeted subpopulations of cells simultaneously, imaging in multiple brain areas concurrently, and superior capabilities for simultaneous Ca²⁺ imaging and optogenetic manipulation. These capacities will further expand the pregnant set of possibilities for optical studies of memory in behaving animals.

ADVANTAGES OF Ca²⁺ IMAGING FOR ANALYSES OF LONG-TERM MEMORY

Beyond its capacities for recording from many individual cells and targeting cells of specific types, both of which are generically useful for systems neuroscience, Ca²⁺ imaging offers additional capabilities that are especially valuable for studies of long-term memory. The latter include the abilities to track individual neurons over many weeks, isolate the signals of neurons with very low activity rates, and visualize the anatomical organization of information processing at cellular resolution. These three facilities have specific import for the learning and memory field.

Long-term memories can persist from time-scales of hours to years. To understand how neural ensembles maintain stored information, it is important to have recording techniques that can monitor individual cells’ dynamics over time spans commensurate with the memory durations of interest. Hence, chronic animal preparations for time-lapse in vivo imaging offer key benefits for studies of long-term memory. Time-lapse fluorescence imaging in live adult mice has provided a potent means for repeatedly inspecting the same neurons and even the same dendritic spines, over periods ranging from weeks to over a year (Grutzendler et al. 2002; Trachtenberg et al. 2002; Zuo et al. 2005; Chen et al. 2008, 2012a; Holtmaat et al. 2009; Yang et al. 2009, 2014b; Cruz-Martin et al. 2010; Fu et al. 2012; Lai et al. 2012; Mostany et al. 2013; Attardo et al. 2015). When this capability is combined with Ca²⁺ imaging, one can track individual neurons’ coding properties over weeks with an ease and efficiency that has been lacking from electrophysiological techniques (Huber et al. 2012; Ziv et al. 2013; Peters et al. 2014).

Traditional electrophysiological recording methods have allowed researchers to track the spiking dynamics of modest numbers of individual cells over days to weeks (Thompson and Best 1990). A majority of such long-term electrophysiological studies has focused on the rodent hippocampus, a crucial brain area for the formation of spatial and episodic memories (Eichenbaum 2000; Buzsaki and Moser 2013). Early reports (Muller et al. 1987; Thompson and Best 1990) of long-term electrical recordings in the hippocampus described longer recording durations than those typical of more recent studies (Lever et al. 2002; Kentros et al. 2004; Muzzio et al. 2009; Mankin et al. 2012; McKenzie et al. 2014). For instance, in rats that repeatedly explored a familiar spatial environment, Thompson and Best (1990) reported that 10 hippocampal neurons maintained stable place fields for 6–153 days. By comparison, recent studies of hippocampal coding have generally tracked greater numbers of cells over briefer durations. For instance, Mankin et al. (2012) reported that they maintained electrical recordings from 30 neurons in three rats over 3 days. Muzzio et al. (2009) followed 65 neurons in 14 mice for 4 days. McKenzie et al. (2014) tracked 38 neurons in four rats over 2 days.

This trend toward more cells and shorter recording durations is probably a reflection of increasingly sophisticated analyses of neural coding and the resulting need for datasets with greater numbers of cells, rising appreciation for the importance of repeatability within individual studies and reproducibility by other labs, and progressive improvements in the statistical rigor of the spike sorting procedures used to identify and track individual cells’ spike waveforms. Crucially, studies based on chronic electrical recordings should include statistical tests confirming that the variations between individual cells’ spike waveforms across consecutive recording days are smaller than the variations between the waveforms of neighboring cells recorded on the same day. Without this, the argument that electrical recordings suffer minimal drift over time is compromised. Overall, the technical challenges inherent to maintaining stable electrical recordings in behaving animals have limited the recording durations and numbers of neurons used for analyses of memory codes. This in turn has constrained the complexity of the questions about long-term memory that researchers have been able to address empirically. Hence, a major advantage of Ca²⁺ imaging is the ability to follow hundreds of individual neurons in a single animal over weeks (Ziv et al. 2013; Peters et al. 2014) yielding datasets that electrical recordings cannot match in statistical power for analyses of neural coding.

In addition to increasing the number of cells whose activity can be tracked in vivo, Ca²⁺ imaging allows researchers to reliably track neurons with very low rates of activity. The ability to monitor such cells is important, because even neurons with very low activity rates can make substantial contributions to information coding if their spikes convey reliable signals. With extracellular electrical recording methods, a certain minimum number of spikes is required to isolate and extract the action potentials from an individual cell by spike sorting, because of the statistical variability of action potential waveforms. This makes it nearly impossible for researchers to track a neuron across 1 or more days of electrical recording in which the neuron fired no spikes. By comparison, in long-term Ca²⁺-imaging studies one tracks individual cells across days by image registration (i.e., by aligning the fluorescence images across imaging sessions). Such image alignment can often be done to <1 µm scale accuracy and does not depend critically on cells’ activity rates (Ziv et al. 2013). Hence, with Ca²⁺ imaging one can track a neuron's dynamics even across long recording epochs in which it is silent. An electrophysiologist experienced with extracellular recordings might argue that this advantage of Ca²⁺ imaging is mainly a theoretical one, because in practical experience neurons are rarely perfectly silent for extended durations. It is important to remember that any such appeal to experience is based on the specific subset of neurons that can be followed electrically and neglects the cells that are too silent for spike sorting.

Indeed, neuroscientists are increasingly realizing that extracellular electrical recordings may often provide a sampling of cells that is biased toward active cells (Lutcke et al. 2013). Studies using intracellular whole-cell patch electrical recordings in both head fixed (Margrie et al. 2002) and freely behaving (Lee et al. 2006) animals support this observation. To achieve whole cell patch electrode recordings the physiologist identifies a neuron from its membrane impedance, not its spiking activity (Margrie et al. 2002). Hence, like Ca²⁺ imaging, these recordings can sample very quiet cells, and the mean neuronal activity rates obtained with the two techniques are broadly consistent. For example, whole cell recording and Ca²⁺-imaging studies of the rodent motor cortex reported spiking and Ca²⁺ transient rates of ∼0.01–1.8 Hz (Lee et al. 2006) and ∼0.001–0.15 Hz (Komiyama et al. 2010; Huber et al. 2012), respectively, one or more orders of magnitude lower than those observed by extracellular recording (∼1–2 Hz) (Laubach et al. 2000). In one Ca²⁺-imaging study of motor learning the rates of motor cortical neuron activation were as low as ∼3.6–0.7 spikes per hour (Huber et al. 2012). Neurons with activity rates this low would be highly challenging to isolate by extracellular electrical recording methods. The degree to which the different methods will yield activity rates out of accord with one another will clearly vary by brain area, cell type, and behavioral state. Regardless of the exact extent of such differences, the ability to follow quiet neurons is especially important for studies of long-term memory, because different subsets of cells may be active on different days (Ziv et al. 2013; Peters et al. 2014).

Recent time-lapse imaging studies of neural ensembles in hippocampus and motor cortex well illustrate Ca²⁺ imaging's methodological advantages. To study the long-term dynamics of CA1 ensemble place codes, Ziv et al. (2013) performed Ca²⁺ imaging in mice exploring a familiar environment and tracked 515–1040 individual CA1 hippocampal pyramidal cells across 45 days in each of four mice (Fig. 5). It was important to verify that when monitored by Ca²⁺ imaging, CA1 neurons displayed coding attributes normally expected of place cells. As anticipated, many neurons exhibited Ca²⁺ activity when the mouse explored a specific portion of its arena. Many of the optically determined place fields remapped when the mouse was transferred to a different arena placed at the same location in the laboratory room, as reported previously (Leutgeb et al. 2005). When the mouse explored a linear track, the set of place fields fully covered the track, and many neurons had statistically significant place fields when the mouse was heading in only one of the two possible running directions, matching past observations (McHugh et al. 1996). About 20% of the cells seen in individual imaging sessions had significant place fields for one or both running directions, and the sizes of the place fields determined by Ca²⁺ imaging were consistent with prior reports (Dombeck et al. 2010; McHugh et al. 1996; Rotenberg et al. 1996; Nakazawa et al. 2003). These observations lent confidence that the pyramidal cells’ Ca²⁺ activation patterns accurately conveyed their place coding properties. By tracking these coding properties as the mice explored the linear track during sessions spaced at 5-day intervals, Ziv et al. (2013) attained some unexpected findings regarding the long-term dynamics of the CA1 representation of a familiar environment.

Figure 5.

Large-scale Ca²⁺ imaging reveals the long-term dynamics of CA1 hippocampal place codes. (A) Using the integrated microscope, Ca²⁺ activity was recorded in 826 CA1 hippocampal pyramidal neurons in a mouse that ran in a linear track during 10 imaging sessions across a 45-day period. Colors indicate the number of sessions in which each cell was active, using the same color scheme as in B. Scale bar, 100 µm. (B) Histogram of the number of sessions in which each of 2960 cells from four mice was active. Error bars are s.d. from counting statistics. (Inset) A constant fraction of all cells detected over 10 sessions was active on each day. Colored data, individual mice; black, mean ± s.e.m. (C) If a cell had Ca²⁺ activity in one session, the odds (blue data) that it also did in a subsequent session declined with time. If a cell had a statistically significant place field in one session, the odds (red data) that it had a place field in a subsequent session also declined with time. Error bars are s.e.m. (D) Distributions of centroid shifts (colored by days between sessions, mean ± s.e.m.) were indistinguishable (Kolmogorov–Smirnov test, P ≥ 0.17), sharply peaked at zero, and highly distinct from the null hypothesis that place fields would randomly relocate (P = 4 × 10⁻⁶⁷, Kolmogorov–Smirnov test). (Inset) Cumulative histograms of shift magnitudes; 74%–83% were ≤ 7 cm. Median shift (3.5 cm) was much less than the median place field width (24 cm). (E) A Bayesian decoding algorithm accurately estimated the mouse's trajectory in the linear track based on the Ca²⁺-imaging data from area CA1 and retained its fidelity across 30 days of experimentation. Reconstructions of the mouse's trajectory (colored curves) and its actual position (black curves); three paired reconstructions comparing time-lapse decoders trained on data from day 5 (right column), using all cells with place fields on both days of each pair, and decoders trained on data from the same day as the test trial (left column). Each pair used an equal number of cells, optimally chosen at left to minimize errors. Scale bars, 2 sec (horizontal) and 10 cm (vertical). (F) Median errors in estimating the mouse's position were ∼7–13 cm, even for decoders trained on data from 30 days prior (black, mean ± s.e.m.). (Red) Decoders trained on data from the same day as test data, using equal numbers of cells as black points and optimally chosen to minimize errors. (Gray) Errors using shuffled traces of Ca²⁺ activity from the same day as training data (averaged over 10,000 shuffles). (G) Cumulative distributions of decoding error magnitudes (mean ± s.e.m.) for test and training data separated by the indicated times or (gray) for decoders tested on shuffled data (all panels are from Ziv et al. 2013).

Notably, there was substantial turnover from session to session in the set of neurons that displayed Ca²⁺ activity (Fig. 5A,B). Only 31 ± 1% (mean ± s.d.) of all neurons seen across the entire experiment were active in any one session, and this percentage was constant over the entire study (Fig. 5B, inset). Across sessions, individual cells came in and out of this active subset in an apparently random manner (Fig. 5A,B), and the overlap in the active subsets from different days declined moderately from ∼60% for sessions 5 days apart to ∼40% for 30 days apart (Fig. 5C). Among the cells that had statistically significant place fields in one or more sessions, the overlap in these coding subsets from different days was ∼25% for sessions 5 days apart and ∼15% for sessions 30 days apart (Fig. 5C). Thus, even in a familiar environment there is substantial dynamism in the CA1 ensemble representation of space. Strikingly, the odds of a cell's recurrence in either the active or coding ensembles were uncorrelated with the rate and amplitude of its Ca²⁺ activity and with the stability of its place field within individual sessions. However, when individual cells did show place fields in more than one session, the place fields’ locations on the linear track were nearly always unchanged (Fig. 5D). Thus, individual cells’ place fields were spatially invariant but temporally stochastic in their day-to-day appearances.

Is ∼15%–25% overlap in the set of cells with significant place fields sufficient to maintain a stable representation of space over weeks? To address this question, Ziv et al. (2013) used computational decoding methods to analyze the spatial information content of the ensemble activity patterns. With datasets containing 515–1040 cells per mouse, it was possible to reconstruct each animal's locomotor trajectory using the representations of space encoded within the ensemble neural dynamics. Despite individual cells’ fluctuating contributions to place coding across different imaging sessions, the ensemble representations of space maintained their coding fidelity for at least a month (Fig. 5D–G) (Ziv et al. 2013). In essence, this fidelity at the ensemble level arose from the spatial invariance of the individual cells’ place fields when they recurred.

Overall, the results of Ziv et al. (2013) suggest quite a different picture of hippocampal spatial coding than had emerged from electrophysiological studies. Certainly, the Ca²⁺ imaging and electrophysiological data are mutually consistent, in that both reveal individual CA1 neurons that stably express place fields over the long term. However, through its capabilities for tracking neurons across sessions in which the neurons are silent, Ca²⁺ imaging reveals that such long-term stability is not the norm. For each location in a familiar environment, it seems there are more than enough cells with corresponding place fields to encode the mouse's position; only a portion of these cells are active in any one session, but this portion is sufficient for stable spatial coding. What is the functional role of the ∼75%–85% of coding cells that do not overlap between any two sessions?

One possibility is that this ∼75%–85% of cells endows each episode with a unique neural signature, potentially allowing the hippocampus to retain distinct memories of the different episodes while preserving information about the shared environment in the other ∼15%–25%. Although speculative, this proposition about episodic memory storage should be experimentally testable, by using long-term Ca²⁺ imaging, decoding analyses, and well-controlled behavioral assays to dissect different cells’ contributions to memories’ spatial and episodic components. By combining Ca²⁺ imaging with molecular manipulations, or with other optical approaches such as fluorescence tagging (Reijmers et al. 2007; Liu et al. 2012; Guenthner et al. 2013; Kawashima et al. 2013; Ramirez et al. 2013; Redondo et al. 2014) and long-term imaging of CA1 dendritic spines (Attardo et al. 2015), it should also be possible to study the mechanisms underlying the spatial stability of the place fields and the temporal variability of their expression.

Another illustration of the substantial statistical analyses that are feasible using large-scale Ca²⁺-imaging data comes from a study of the mouse motor cortex that tracked ∼200 neurons in each of 10 mice as the animals learned a lever-press task over 14 days (Fig. 6A–G) (Peters et al. 2014). The data revealed substantial reorganization and refinement of the ensemble neural dynamics over the course of motor learning. During the first 3 days, the number of neurons with movement-related activity patterns approximately doubled, from ∼10% to ∼20%; across subsequent days, the number of such cells gradually returned to near its initial value (Fig. 6A). Throughout training, the timing of neural responses relative to movement onset gradually became less variable (Fig. 6B) and more correlated across movement trials (Fig. 6C–E).

Figure 6.

Large-scale Ca²⁺ imaging reveals the reorganization of ensemble neural dynamics in mouse layer 2/3 motor cortical neurons across a 2-week regimen of daily motor learning. (A) Fraction of excitatory neurons classified as movement related in each of the 14 training sessions, during which mice learned a lever-press task. (B) Standard deviation of the timing of activity onsets for movement-related excitatory neurons for neurons that were active in five or more sessions. The trial-by-trial variability in the onset of single cell activity declined as learning progressed. (C) Ensemble neural activity became more stereotyped as learning advanced, as shown by the increase in the pairwise trial-to-trial correlation of temporal population activity vectors. The temporal population activity vector was a concatenation of the activity traces of all movement-related neurons and thus maintained temporal information within each movement. (D) Neuronal activity increased in temporal structure over the course of motor learning. Raster plots of neuronal Ca²⁺ activity for movement-related neurons that were active in at least 10% of trials on the sessions indicated, shown for an example animal. Red arrows indicate movement onset. Colors distinguish individual neurons that have been sorted according to their preferred timing. Note that the same color across different sessions does not necessarily mark the same neurons. (E) Maximum-normalized average activity from all movement-related neurons from all animals in session 2 (106 neurons) and session 14 (84 neurons) aligned to movement onset (red arrow). The timing of neuronal activity was more refined in session 14, as evidenced by the narrower peaks and lower levels of background activity, and the timing shifts toward movement onset. (F) In expert but not naïve mice, the extent to which ensemble neural activity in individual trials correlated with the average learned neural activity pattern from expert sessions was predictive of the extent to which the animal's movement pattern correlated with the learned movement pattern. In sessions during which the mice were naïve, when the animals made movements that resembled the learned movement patterns acquired later in training, the underlying neural activity patterns were nevertheless very different from the learned neural activity pattern late in training. (G) Pairwise trial-to-trial correlation of temporal population neural activity vectors (as in C) plotted as a function of movement correlation on those trials. A stronger relationship between ensemble neural activity and movement emerges during learning. All error bars are s.e.m. (all panels are modified from Peters et al. 2014).

Moreover, as the mice became expert and learned a stereotyped movement, the neural ensembles gradually acquired stereotyped activity patterns, and the degree to which the ensemble dynamics on individual trials resembled these stereotyped patterns was predictive of the degree to which the animal's movement resembled its expert, stereotyped form (Fig. 6F,G). However, even in mice that were expert at performing stereotyped movements, there remained substantial trial-to-trial variability in neurons’ responses (Fig. 6F,G). Like the variability in CA1 place cells’ dynamics that left ensemble spatial coding fidelity largely intact, this finding suggests that stable retention of information at the level of neural populations does not require purely stable neuronal firing properties. On the contrary, neural ensembles may be far more robust in their information-processing capabilities than can be inferred from the activity patterns of individual cells.

Notably, though Ca²⁺ imaging affords access to the topographic organization of memory at a cellular scale, the recent imaging studies of both hippocampus and motor cortex did not find any clear anatomical patterns by which cells with similar coding properties are organized (Dombeck et al. 2010; Ziv et al. 2013; Peters et al. 2014). Instead, the arrangements appeared to be random. Similarly, a recent study of piriform cortex yielded evidence that cells involved in storing learned odor memories are also randomly organized (Choi et al. 2011). By comparison, a recent Ca²⁺-imaging study of medial entorhinal cortex found a clear micro-organization in the anatomical arrangements of grid cells (Heys et al. 2014). In sensory and motor brain areas there are well-known topographic maps that relate neurons’ anatomic locations and coding properties to specific features of the sensory world or the animal's body plan (Albright et al. 2000). Particularly in cognitive but also in motor areas, the manner in which information processing is organized at the cellular scale remains far less explored. Plainly, large-scale in vivo Ca²⁺ imaging provides a potent new tool to do so, for multiple forms of information processing (Dombeck et al. 2010; Ziv et al. 2013; Heys et al. 2014; Peters et al. 2014).

Overall, Ca²⁺ imaging permits memory researchers to probe aspects of neural coding that have been inaccessible from electrical recordings. The sheer number of neurons that neuroscientists can now reliably image over weeks enables wholly new analyses of how ensemble codes form, evolve, and are refined over the course of memory encoding, consolidation, maintenance, and recall. As individual neurons enter, exit, or return to the ensemble codes involved in memory processing, neuroscientists can track these cellular transitions, even over extended time periods during which some of the individual cells are silent. By using Ca²⁺ imaging, memory researchers can also examine how cells’ anatomical arrangements relate to their memory processing roles. Each of these new capacities will likely provide important insights regarding the brain's rules and mechanisms for handling memory data.

PROSPECTS FOR STUDYING THE ORGANIZATION OF MEMORY USING Ca²⁺ IMAGING

Memories are often linked to one another, such as through associations and shared attributes, and organized into structures such as categories, sequences, and hierarchies. By providing the means to examine the ensemble neural codes supporting two or more different memories, in vivo Ca²⁺ imaging offers intriguing prospects for directly visualizing the biological instantiation of memory organization. The large numbers of neurons that can be monitored currently by Ca²⁺ imaging will likely be crucial for this pursuit, because substantial statistical power will be necessary to determine the extent to which related memories are supported by overlapping neural representations, and whether these representations use local or distributed coding schemes. Moreover, the brain's memory representations are generally dynamic and involve multiple brain regions to varying extents at different stages of memory processing. As a memory becomes more remote in time, the reshaping of its representation across different brain areas is known as systems memory consolidation (Dudai 2004; Frankland and Bontempi 2005). Time-lapse Ca²⁺-imaging studies may provide a window onto these events at the cellular scale in live animals.

A large body of research in psychology and cognitive neuroscience has focused on the extent to which memory representations are local or distributed, and how such representations encode the relationships between different memories. There is a spectrum of theoretical possibilities (Quian Quiroga and Kreiman 2010). In theories at the “localist” end of the spectrum (Page 2000; Bowers 2009), individual memories reside within individual cells; cell identities provide individual memories their unique codes or signatures and synaptic connections link different memories together. By comparison, in theories involving fully distributed representations, individual cells are generally involved in the encoding of multiple memories (Rumelhart et al. 1986; Masson 1987; Plaut and McClelland 2010); in these theories, it is the distributed patterns of neural activity that provide unique signatures to individual memories, hence, memories can only be deciphered from the population-level activity patterns and not the dynamics of individual neurons. Clearly, there are many intermediate possibilities as well, melding these aspects of representation for different facets of memory and across different brain structures. Notably, the vast majority of such ideas from cognitive science have not yet been tested empirically through biological studies of neural dynamics. In vivo Ca²⁺-imaging studies might change this situation, by enabling direct observations of the extent to which cell identities or activity patterns provide memories their signatures—at least in ensembles of cells that can be imaged simultaneously.

In addition to the newfound experimental possibilities for examining whether memory representations are local or distributed, Ca²⁺ imaging also offers new means to examine the manner in which the brain encodes associations, which can link different parts of an individual memory and relate different memories to each other. Consider an example Ca²⁺-imaging study of second-order associative conditioning. Presently, there are ongoing Ca²⁺-imaging studies of first-order associative learning in brain structures such as the amygdala and nucleus accumbens (Grewe et al. 2014; Parker et al. 2014), and it is reasonable to envision these experiments could be suitably extended. In second-order conditioning (Gewirtz and Davis 2000; Debiec et al. 2006), an animal first experiences paired presentations of an initially neutral conditioned stimulus (CS₁) and an unconditioned stimulus (US) that evokes an unconditioned behavioral response. With learning, the animal gains a conditioned response to CS₁. A second conditioned stimulus (CS₂) is then paired with CS₁; afterward, a presentation of CS₂ evokes the conditioned response despite that CS₂ was never directly paired with the US. How is the relationship between the US and CS₂ encoded?

One proposal is that the individual representations of CS₁ and CS₂ involve overlapping populations of neurons (Debiec et al. 2006). However, any such overlap has not been directly observed, in part because of the long-standing inability to visualize ensemble neural activity throughout learning and memory formation. By using in vivo Ca²⁺ imaging in a brain area that has been implicated in forming the associations for a particular form of learning, it may be feasible to directly examine the neural codes involved. Specifically, it should be possible to visualize the extent to which the cellular representations of CS₂, CS₁, and the US overlap, as determined from the subsets of neurons activated by each stimulus. Of key interest is the manner in which the representation of the association between CS₂ and the US depends on the prior association between CS₁ and the US. Will the associative relationships be encoded by overlapping cellular memberships in the representations (Debiec et al. 2006), distinctive signatures in the neural ensemble activity (Buzsaki 2005; Buzsaki and Moser 2013), combinations thereof, or means that are not observable by Ca²⁺ imaging? More generally, a second-order association is only one type of relationship between memories, and Ca²⁺ imaging might help elucidate the organization of neural codes and their interrelationships for multiple other forms of learning. For example, categorical memories may arise from more elementary ones, and future Ca²⁺-imaging studies might allow sophisticated analyses of this process.

A related set of issues concerns the temporal relationships between different memories and the manner in which new information is encoded in relation to prior memories. The biological mechanisms that govern the encoding of temporal relationships are only starting to be uncovered. Notably, neurons expressing virally induced, elevated levels of the cAMP-response element-binding protein (CREB) appear to have increased electrical excitability and may be primed for upcoming bouts of memory storage (Han et al. 2007, 2009; Zhou et al. 2009; Yiu et al. 2014). Further, neurons that have recently encoded a memory, as assessed by CREB phosphorylation, maintain their elevated levels of phosphorylated CREB and presumed heightened excitability for several hours following learning; this persistence of CREB activation may bias the storage of subsequent memories toward the same set of excitable neurons encoding the first memory (Silva et al. 2009). Thus, two learning episodes that are temporally separated, but within the time window of CREB activation, may be encoded in overlapping populations of neurons, thereby forming an association between the two memories in the brain (Rogerson et al. 2014). Similarly, behavioral (Ballarini et al. 2009) and cell biological (Frey and Morris 1997) phenomena in which one set of events impacts the encoding of other temporally separated events indicate that memories are not encoded in isolation, but rather in relation to prior and subsequent events (Moncada and Viola 2007). To investigate issues of this kind, multiple studies to date have examined memory representations via patterns of IEG activation (Guzowski et al. 1999; Frankland and Bontempi 2005; Reijmers et al. 2007; Goshen et al. 2011; Tse et al. 2011; Czajkowski et al. 2014), but Ca²⁺-imaging datasets might offer richer possibilities for analyses of the spatiotemporal features of ensemble neural coding.

Another important aspect of memory organization that may be ripe for study by Ca²⁺ imaging is systems memory consolidation (Dudai 2004; Frankland and Bontempi 2005). Substantial literature suggests that as a stored memory becomes more remote, neocortical memory systems may gradually extract and reorganize memory information from an initially acting memory system such as hippocampus (McClelland et al. 1995), eventually obviating the necessity of hippocampus for memory recall (Scoville and Milner 1957; Kim and Fanselow 1992; Squire and Alvarez 1995; Wiltgen et al. 2004; Bontempi and Durkin 2007; Tse et al. 2007, 2011). These rearrangements of memories across brain areas might allow the neocortex to organize information in a more abstract form (Fuster 2009), such as by using schemas that establish higher-order associations and categories of information or knowledge (Piaget et al. 1929; Bartlett 1932; van Kesteren et al. 2012; McKenzie et al. 2013; Preston and Eichenbaum 2013).

However, because of the extended time course of systems consolidation, from weeks in rodents (Kim and Fanselow 1992) to years in humans (Scoville and Milner 1957; Squire and Alvarez 1995), the relevant interactions between neural ensembles in hippocampus and neocortex remain unknown. By using time-lapse Ca²⁺-imaging ability to track neuronal ensembles over extended durations, researchers might conceivably be able to watch how large sets of individual neurons in different brain regions exchange information. Optical methods for imaging multiple brain areas concurrently (Lecoq et al. 2014), including hippocampus and neocortex, may be especially useful.

Recent studies in rodents have also challenged the long-held notions that the engagement of neocortex in systems consolidation is always slow (Tse et al. 2007; Tse et al. 2011), and that the necessity of hippocampus for memory recall always diminishes over the course of long-term memory storage (Wiltgen et al. 2010; Goshen et al. 2011). By using IEG expression as a readout of cellular involvement in memory formation, studies in rats have shown that once a schema is established, the neocortex can be directly activated by new information that fits in the schema, in a necessary manner for learning (Tse et al. 2007; Tse et al. 2011). Recent studies in mice indicate a sustained role for hippocampus in the activation of neocortical circuits (Goshen et al. 2011) and in the recall of detailed memories (Wiltgen et al. 2010) at 2–4 weeks after memory storage. Overall, the anatomical and time-dependent reorganization of memory representations following their initial formation deserves intensive further study; Ca²⁺ imaging may allow a much richer set of observations regarding how schemas are established and impact the subsequent encoding of new information and the engagement of multiple brain areas. Other brain structures beyond hippocampus and neocortex also seem to undergo processes similar to systems consolidation (Repa et al. 2001; Do-Monte et al. 2015), indicating the general need for methods capable of revealing how the brain reorganizes memories over time.

OUTLOOK

The advent of powerful technologies for large-scale Ca²⁺ imaging in behaving mammals has provided major new opportunities to study the ensemble neural dynamics that underlie learning and memory. We expect active advancements on multiple research fronts that will continue to improve the potency of Ca²⁺ imaging for such studies. Areas of technical progress will almost certainly include advances in new Ca²⁺ indicators (Fosque et al. 2015; Inoue et al. 2015), viral and genetic labeling methods (Oyibo et al. 2014; Madisen et al. 2015), optical instrumentation for imaging in one or more brain areas (Lecoq et al. 2014; Stirman et al. 2014), the combination of Ca²⁺ imaging and optogenetics (Grosenick et al. 2015), behavioral assays expressly designed or optimized for Ca²⁺-imaging studies (Lovett-Barron et al. 2014; Sofroniew et al. 2014), and computational analyses of large-scale neural-imaging data (Ziv et al. 2013; Peters et al. 2014). By combining such technologies, memory researchers will have unprecedented capabilities to decipher the ensemble neural codes that support long-term memory across a range of animal behaviors and brain areas, and to dissect the rules that govern memory storage. To bring the suite of available techniques to full fruition, collaborations between scientists with diverse expertise will become ever more crucial, as cutting-edge experiments increasingly require multidisciplinary knowledge in fields such as genetics, virology, cognitive and systems neuroscience, optics, and computer science. However, if the research community can effectively harness its collective know-how, we stand to gain important insights into long-standing questions about the organization of memory in the mammalian brain.